Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,306 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# ---------------------------------------------------------------------------------------
|
2 |
+
# Imports and Options
|
3 |
+
# ---------------------------------------------------------------------------------------
|
4 |
+
import streamlit as st
|
5 |
+
import pandas as pd
|
6 |
+
import requests
|
7 |
+
import re
|
8 |
+
import fitz # PyMuPDF
|
9 |
+
import io
|
10 |
+
import matplotlib.pyplot as plt
|
11 |
+
from PIL import Image
|
12 |
+
from mlx_vlm import load, generate
|
13 |
+
from mlx_vlm.prompt_utils import apply_chat_template
|
14 |
+
from mlx_vlm.utils import load_config, stream_generate
|
15 |
+
from docling_core.types.doc.document import DocTagsDocument, DoclingDocument
|
16 |
+
|
17 |
+
# Set Streamlit to wide mode
|
18 |
+
# st.set_page_config(layout="wide")
|
19 |
+
|
20 |
+
# ---------------------------------------------------------------------------------------
|
21 |
+
# API Configuration
|
22 |
+
# ---------------------------------------------------------------------------------------
|
23 |
+
API_URL = "https://api.stack-ai.com/inference/v0/run/2df89a6c-a4af-4576-880e-27058e498f02/67acad8b0603ba4631db38e7"
|
24 |
+
headers = {
|
25 |
+
'Authorization': 'Bearer a9e4979e-cdbe-49ea-a193-53562a784805',
|
26 |
+
'Content-Type': 'application/json'
|
27 |
+
}
|
28 |
+
|
29 |
+
# ---------------------------------------------------------------------------------------
|
30 |
+
# Survey Analysis Class
|
31 |
+
# ---------------------------------------------------------------------------------------
|
32 |
+
class SurveyAnalysis:
|
33 |
+
def __init__(self, api_key=None):
|
34 |
+
self.api_key = api_key
|
35 |
+
|
36 |
+
def prepare_llm_input(self, survey_response, topics):
|
37 |
+
# Create topic description string from user input
|
38 |
+
topic_descriptions = "\n".join([f"- **{topic}**: {description}" for topic, description in topics.items()])
|
39 |
+
|
40 |
+
llm_input = f"""
|
41 |
+
Your task is to review PDF docling and extract information related to the provided topics. Here are the topic descriptions:
|
42 |
+
|
43 |
+
{topic_descriptions}
|
44 |
+
|
45 |
+
**Instructions:**
|
46 |
+
- Extract and summarize the PDF focusing only on the provided topics.
|
47 |
+
- If a topic is not mentioned in the notes, it should not be included in the Topic_Summary.
|
48 |
+
- Use **exact quotes** from the original text for each point in your Topic_Summary.
|
49 |
+
- Exclude erroneous content.
|
50 |
+
- Do not add additional explanations or instructions.
|
51 |
+
|
52 |
+
**Format your response as follows:**
|
53 |
+
[Topic]
|
54 |
+
- "Exact quote"
|
55 |
+
- "Exact quote"
|
56 |
+
- "Exact quote"
|
57 |
+
|
58 |
+
**Meeting Notes:**
|
59 |
+
{survey_response}
|
60 |
+
"""
|
61 |
+
return llm_input
|
62 |
+
|
63 |
+
def query_api(self, payload):
|
64 |
+
response = requests.post(API_URL, headers=headers, json=payload)
|
65 |
+
return response.json()
|
66 |
+
|
67 |
+
def extract_meeting_notes(self, response):
|
68 |
+
output = response.get('outputs', {}).get('out-0', '')
|
69 |
+
return output
|
70 |
+
|
71 |
+
def process_dataframe(self, df, topics):
|
72 |
+
results = []
|
73 |
+
for _, row in df.iterrows():
|
74 |
+
llm_input = self.prepare_llm_input(row['Document_Text'], topics)
|
75 |
+
payload = {
|
76 |
+
"user_id": "<USER or Conversation ID>",
|
77 |
+
"in-0": llm_input
|
78 |
+
}
|
79 |
+
response = self.query_api(payload)
|
80 |
+
meeting_notes = self.extract_meeting_notes(response)
|
81 |
+
results.append({
|
82 |
+
'Document_Text': row['Document_Text'],
|
83 |
+
'Topic_Summary': meeting_notes
|
84 |
+
})
|
85 |
+
|
86 |
+
result_df = pd.DataFrame(results)
|
87 |
+
df = df.reset_index(drop=True)
|
88 |
+
return pd.concat([df, result_df[['Topic_Summary']]], axis=1)
|
89 |
+
|
90 |
+
# ---------------------------------------------------------------------------------------
|
91 |
+
# Function to Extract Excerpts
|
92 |
+
# ---------------------------------------------------------------------------------------
|
93 |
+
def extract_excerpts(processed_df):
|
94 |
+
new_rows = []
|
95 |
+
|
96 |
+
for _, row in processed_df.iterrows():
|
97 |
+
Topic_Summary = row['Topic_Summary']
|
98 |
+
|
99 |
+
# Split the Topic_Summary by topic
|
100 |
+
sections = re.split(r'\n(?=\[)', Topic_Summary)
|
101 |
+
|
102 |
+
for section in sections:
|
103 |
+
# Extract the topic
|
104 |
+
topic_match = re.match(r'\[([^\]]+)\]', section)
|
105 |
+
if topic_match:
|
106 |
+
topic = topic_match.group(1)
|
107 |
+
|
108 |
+
# Extract all excerpts within the section
|
109 |
+
excerpts = re.findall(r'- "([^"]+)"', section)
|
110 |
+
|
111 |
+
for excerpt in excerpts:
|
112 |
+
new_rows.append({
|
113 |
+
'Document_Text': row['Document_Text'],
|
114 |
+
'Topic_Summary': row['Topic_Summary'],
|
115 |
+
'Excerpt': excerpt,
|
116 |
+
'Topic': topic
|
117 |
+
})
|
118 |
+
|
119 |
+
return pd.DataFrame(new_rows)
|
120 |
+
|
121 |
+
#------------------------------------------------------------------------
|
122 |
+
# Streamlit Configuration
|
123 |
+
#------------------------------------------------------------------------
|
124 |
+
|
125 |
+
# Set page configuration
|
126 |
+
st.set_page_config(
|
127 |
+
page_title="Choose Your Own Adventure (Topic Extraction) PDF Analysis App",
|
128 |
+
page_icon=":bar_chart:",
|
129 |
+
layout="centered",
|
130 |
+
initial_sidebar_state="auto",
|
131 |
+
menu_items={
|
132 |
+
'Get Help': 'mailto:[email protected]',
|
133 |
+
'About': "This app is built to support PDF analysis"
|
134 |
+
}
|
135 |
+
)
|
136 |
+
|
137 |
+
#------------------------------------------------------------------------
|
138 |
+
# Sidebar
|
139 |
+
#------------------------------------------------------------------------
|
140 |
+
|
141 |
+
# Sidebar with image
|
142 |
+
with st.sidebar:
|
143 |
+
# Set the desired width in pixels
|
144 |
+
image_width = 300
|
145 |
+
# Define the path to the image
|
146 |
+
# image_path = "steelcase_small.png"
|
147 |
+
image_path = "/Users/clevesse/Documents/VSC_Code/PDF_Extraction/PDF_Extraction_streamlit/steelcase_small.png"
|
148 |
+
# Display the image
|
149 |
+
st.image(image_path, width=image_width)
|
150 |
+
|
151 |
+
# Additional sidebar content
|
152 |
+
|
153 |
+
with st.expander("**WorkSpace Futures**", expanded=True):
|
154 |
+
st.write("""
|
155 |
+
Strategic Market Intelligence
|
156 |
+
Director: Amy Willard
|
157 |
+
|
158 |
+
- **Support**: Cheyne LeVesseur PhD
|
159 |
+
- **Email**: [email protected]
|
160 |
+
""")
|
161 |
+
st.divider()
|
162 |
+
st.subheader('Instructions')
|
163 |
+
|
164 |
+
Instructions = """
|
165 |
+
- **Step 1**: Upload your PDF file.
|
166 |
+
- **Step 2**: Review the processed meeting notes with extracted excerpts and classifications.
|
167 |
+
- **Step 3**: Review topic descriptions.
|
168 |
+
- **Step 4**: Review topic distribution and frequency.
|
169 |
+
- **Step 5**: Review bar charts of topics.
|
170 |
+
- **Step 6**: Download the processed data as a CSV file.
|
171 |
+
"""
|
172 |
+
st.markdown(Instructions)
|
173 |
+
|
174 |
+
# Load SmolDocling model (mlx_vlm version)
|
175 |
+
@st.cache_resource
|
176 |
+
def load_smol_docling():
|
177 |
+
model_path = "ds4sd/SmolDocling-256M-preview-mlx-bf16"
|
178 |
+
model, processor = load(model_path)
|
179 |
+
config = load_config(model_path)
|
180 |
+
return model, processor, config
|
181 |
+
|
182 |
+
model, processor, config = load_smol_docling()
|
183 |
+
|
184 |
+
# Convert PDF to images
|
185 |
+
def convert_pdf_to_images(pdf_file):
|
186 |
+
images = []
|
187 |
+
doc = fitz.open(stream=pdf_file.read(), filetype="pdf")
|
188 |
+
for page_number in range(len(doc)):
|
189 |
+
page = doc.load_page(page_number)
|
190 |
+
pix = page.get_pixmap(dpi=300) # Higher DPI for clarity
|
191 |
+
img_data = pix.tobytes("png")
|
192 |
+
image = Image.open(io.BytesIO(img_data))
|
193 |
+
images.append(image)
|
194 |
+
return images
|
195 |
+
|
196 |
+
# Extract structured markdown text using SmolDocling (mlx_vlm)
|
197 |
+
def extract_markdown_from_image(image):
|
198 |
+
prompt = "Convert this page to docling."
|
199 |
+
formatted_prompt = apply_chat_template(processor, config, prompt, num_images=1)
|
200 |
+
output = ""
|
201 |
+
|
202 |
+
for token in stream_generate(
|
203 |
+
model, processor, formatted_prompt, [image], max_tokens=4096, verbose=False):
|
204 |
+
output += token.text
|
205 |
+
if "</doctag>" in token.text:
|
206 |
+
break
|
207 |
+
|
208 |
+
# Convert DocTags to Markdown
|
209 |
+
doctags_doc = DocTagsDocument.from_doctags_and_image_pairs([output], [image])
|
210 |
+
doc = DoclingDocument(name="ExtractedDocument")
|
211 |
+
doc.load_from_doctags(doctags_doc)
|
212 |
+
markdown_text = doc.export_to_markdown()
|
213 |
+
return markdown_text
|
214 |
+
|
215 |
+
# Streamlit UI
|
216 |
+
st.title("Choose Your Own Adventure (Topic Extraction) PDF Analysis App")
|
217 |
+
|
218 |
+
uploaded_file = st.file_uploader("Upload PDF file", type=["pdf"])
|
219 |
+
|
220 |
+
if uploaded_file:
|
221 |
+
with st.spinner("Processing PDF..."):
|
222 |
+
images = convert_pdf_to_images(uploaded_file)
|
223 |
+
|
224 |
+
markdown_texts = []
|
225 |
+
for idx, image in enumerate(images):
|
226 |
+
markdown_text = extract_markdown_from_image(image)
|
227 |
+
markdown_texts.append(markdown_text)
|
228 |
+
|
229 |
+
df = pd.DataFrame({'Document_Text': markdown_texts})
|
230 |
+
|
231 |
+
st.success("PDF processed successfully!")
|
232 |
+
|
233 |
+
# Check if extraction was successful
|
234 |
+
if df.empty or df['Document_Text'].isnull().all():
|
235 |
+
st.error("No meaningful text extracted from the PDF.")
|
236 |
+
st.stop()
|
237 |
+
|
238 |
+
st.markdown("### Extracted Markdown Preview")
|
239 |
+
st.write(df.head())
|
240 |
+
|
241 |
+
# ---------------------------------------------------------------------------------------
|
242 |
+
# User Input for Topics
|
243 |
+
# ---------------------------------------------------------------------------------------
|
244 |
+
st.markdown("### Enter Topics and Descriptions")
|
245 |
+
num_topics = st.number_input("Number of topics", min_value=1, max_value=10, value=1, step=1)
|
246 |
+
|
247 |
+
topics = {}
|
248 |
+
for i in range(num_topics):
|
249 |
+
topic = st.text_input(f"Topic {i+1} Name", key=f"topic_{i}")
|
250 |
+
description = st.text_area(f"Topic {i+1} Description", key=f"description_{i}")
|
251 |
+
if topic and description:
|
252 |
+
topics[topic] = description
|
253 |
+
|
254 |
+
# Add a button to execute the analysis
|
255 |
+
if st.button("Run Analysis"):
|
256 |
+
if not topics:
|
257 |
+
st.warning("Please enter at least one topic and description.")
|
258 |
+
st.stop()
|
259 |
+
|
260 |
+
# ---------------------------------------------------------------------------------------
|
261 |
+
# Your existing SurveyAnalysis and extract_excerpts functions remain unchanged here:
|
262 |
+
# ---------------------------------------------------------------------------------------
|
263 |
+
analyzer = SurveyAnalysis()
|
264 |
+
processed_df = analyzer.process_dataframe(df, topics)
|
265 |
+
df_VIP_extracted = extract_excerpts(processed_df)
|
266 |
+
|
267 |
+
required_columns = ['Document_Text', 'Topic_Summary', 'Excerpt', 'Topic']
|
268 |
+
missing_columns = [col for col in required_columns if col not in df_VIP_extracted.columns]
|
269 |
+
|
270 |
+
if missing_columns:
|
271 |
+
st.error(f"Missing columns after processing: {missing_columns}")
|
272 |
+
st.stop()
|
273 |
+
|
274 |
+
df_VIP_extracted = df_VIP_extracted[required_columns]
|
275 |
+
|
276 |
+
st.markdown("### Processed Meeting Notes")
|
277 |
+
st.dataframe(df_VIP_extracted)
|
278 |
+
|
279 |
+
st.write(f"**Number of meeting notes analyzed:** {len(df)}")
|
280 |
+
st.write(f"**Number of excerpts extracted:** {len(df_VIP_extracted)}")
|
281 |
+
|
282 |
+
# CSV download
|
283 |
+
csv = df_VIP_extracted.to_csv(index=False)
|
284 |
+
st.download_button(
|
285 |
+
"Download data as CSV",
|
286 |
+
data=csv,
|
287 |
+
file_name='extracted_meeting_notes.csv',
|
288 |
+
mime='text/csv'
|
289 |
+
)
|
290 |
+
|
291 |
+
# Topic distribution visualization
|
292 |
+
topic_counts = df_VIP_extracted['Topic'].value_counts()
|
293 |
+
frequency_table = pd.DataFrame({'Topic': topic_counts.index, 'Count': topic_counts.values})
|
294 |
+
frequency_table['Percentage'] = (frequency_table['Count'] / frequency_table['Count'].sum() * 100).round(0)
|
295 |
+
|
296 |
+
st.markdown("### Topic Distribution")
|
297 |
+
st.dataframe(frequency_table)
|
298 |
+
|
299 |
+
fig, ax = plt.subplots(figsize=(10, 5))
|
300 |
+
ax.bar(frequency_table['Topic'], frequency_table['Count'], color='#3d9aa1')
|
301 |
+
ax.set_ylabel('Count')
|
302 |
+
ax.set_title('Frequency of Topics')
|
303 |
+
st.pyplot(fig)
|
304 |
+
|
305 |
+
else:
|
306 |
+
st.info("Please upload a PDF file to begin.")
|